Abstract
<title>Abstract</title> <p>Digital breast tomosynthesis (DBT) is an established imaging modality for breast cancer screening that reduces tissue overlap compared with conventional digital mammography. Publicly available DBT datasets remain scarce, and existing resources predominantly provide coarse image-level labels or 2D bounding boxes. Here, we present BCSDBT-Seg, an open-access, expert-annotated dataset providing 3D voxel-level lesion segmentation masks and structured clinical reports for the publicly available Breast Cancer Screening Digital Breast Tomosynthesis (BCS-DBT) collection. Radiologists manually delineated lesion boundaries on 396 DBT volumes (craniocaudal and mediolateral oblique views) from 201 biopsy-confirmed patients, yielding 434 distinct lesion masks. In addition, structured Breast Imaging Reporting and Data System (BI-RADS) descriptors, breast density ratings, and pathology outcomes are provided per case. Technical validation using 3D nnU-Net baseline models demonstrates the quality and utility of the annotations. BCSDBT-Seg supports applications in automated segmentation, radiomics, and deep learning model development for breast image analysis.</p>